Question: What is the runtime of Fast-FoundationStereo on a 5090?

[sbirchfield]

Topic

Validation, benchmarking, or qualification

Question

What is the runtime of Fast-FoundationStereo on a 5090?

What have you tried?

No response

TensorRT-Model-Connect version

No response

Model and revision

No response

Environment details

No response

Submission checks

• I searched the documentation and open and closed issues first.
• This question contains no secrets, private URLs, or restricted artifacts.

[chaofengw-nv]

Hi @sbirchfield，
Fast-FoundationStereo was validated on RTX 5090 using the implementation in PR #1071.

RTX 5090 Performance

• TensorRT p50 latency: 17.428 ms
• PyTorch reference p50 latency: 36.270 ms
• Reference/candidate latency ratio: 2.081x
• Batch size: 1
• Input resolution: 700 × 700
• Precision: FP16
• max_disp: 192
• valid_iters: 8
• Warmup iterations: 3
• Measured iterations: 10

The tested input was the Fast-FoundationStereo office stereo pair.

Output Validation

The TensorRT output passed the disparity parity checks against the PyTorch reference:

• Cosine similarity: 0.999842
• Mean absolute error: 0.165125
• Bad-2px ratio: 0.018316

Middlebury-Q accuracy validation also passed all 15/15 scenes:

• Candidate/reference EPE: 0.442262 / 0.440257
• Candidate/reference BP2: 0.029331 / 0.028795

Environment

• GPU: NVIDIA GeForce RTX 5090
• TensorRT builder: 11.1.0.106
• TensorRT runtime ABI: 11.1
• CUDA: 13.0
• PyTorch: 2.12.0+cu130
• CUDA architecture: SM120
• Native plugin build option: CMAKE_CUDA_ARCHITECTURES=120-real
• Tested revision: 359905bb533963e5711b01187d29a185bd53f69d

GB300-1 Reference Result

Using the same workload configuration:

• TensorRT: 11.1.0.106
• TensorRT p50 latency: 12.723 ms
• PyTorch reference p50 latency: 52.010 ms
• Reference/candidate latency ratio: 4.088x
• Tested revision: 06b396d63ad00f89ce4f07eb38b373e1a2657145

PR: #1071